Monday, August 14, 2017

Awesome Tag final F1 score 2017.08.14

I returned to the Awesome Tag project last month. I implemented some new tests that improved my F1 score from 0.44 in January to about 0.66 in July. This project is a facial detection scheme. Of course it's not as ambitious as facial recognition might be. On top of that the process of scanning a particular jpeg or png takes several minutes. The code is sloppy. I wanted to reduce the number of false posotives and was somewhat successfull, but the outcome of my efforts is that the F1 score has dropped. Today I calculated my F1 roughly in the 0.11 ballpark. I am getting far fewer true posotives.

The python code from this project uses google's tensorflow to manipulate several neural network type objects on my computer's gpu. At one point I went so far as to write my own gpu tensorflow kernel for doing some of the work involved in testing images for skin tones. Still, the code does not work well and I am leaving the project as it is. I would love to return to the project and somehow achieve good results (and quickly) but I do not think that will happen.

As stated above the F1 score at this point is 0.11, not very good.

No comments:

Post a Comment